Does AI Customer Service Really Work in Arabic in 2026?

plum in practice 6 September 2026 6 min read

A hand holding a smartphone above a plain white desk

Photo by freestocks.org on Flickr, CC0 1.0

A customer in Jeddah asks how much delivery costs, typed in Arabic script, and gets back a polite paragraph of Modern Standard Arabic that reads like a government circular. Another writes "3andkon delivery la Beirut?" in Latin letters and gets nothing useful at all. If you sell across the Gulf and the Levant, this is the daily reality of Arabic customer service, and it is the part most automation demos quietly skip.

Arabic AI customer service works, but not the way English does, and the difference is not translation quality. Arabic customers do not write in one language. They write in Modern Standard Arabic, in their own regional dialect, in Latin-script Arabizi, and in code-switched Arabic and English, often inside a single message. A system that only handles Modern Standard Arabic will answer formal questions correctly and miss most real ones. On WhatsApp there is a second, separate problem: Meta requires you to assign a language code when you create a message template, and Meta does not translate template strings or variables for you. You supply every language yourself. So Arabic support is really two jobs, understanding messy input in the conversation, and maintaining a parallel set of approved Arabic templates for anything you send first. Businesses that do only the first half still sound foreign the moment they initiate.

The input problem is dialect, not translation

Machine translation of Arabic has been decent for years. That is not what breaks.

What breaks is that "Arabic" is not one written register in customer messages. A shopper in Kuwait, a shopper in Casablanca and a shopper in Beirut asking the same question about delivery will produce three visibly different strings, and none of them will look like the Modern Standard Arabic in a textbook. Add Arabizi, where Arabic is typed in Latin characters with numerals standing in for sounds that have no Latin equivalent, and you have a fourth form that many systems treat as gibberish because it is neither Arabic script nor an English word.

Then there is code-switching. A Gulf customer who opens in Arabic script to say the order has not arrived, then finishes the same sentence with "can you check please", is not being unusual. That is a normal sentence.

So the real question to ask a vendor is not "do you support Arabic". Every vendor says yes. The question is what happens to a code-switched Arabizi message with a spelling mistake in it, because that is what actually arrives on a Tuesday afternoon.

How to test it honestly

You can evaluate this in an hour with no technical work. Collect twenty real customer messages from your own WhatsApp history, unedited, keeping the typos and the mixed script. Make sure the set includes:

At least five in dialect rather than Modern Standard Arabic
At least three in Arabizi
At least three that switch between Arabic and English mid-sentence
Two with a genuine spelling mistake
Two asking about something you do not sell, to see how it handles a miss

Then run them and read the replies as a customer, not as the person who bought the tool. A system that gets fifteen of twenty right on your own messages is worth more than any demo.

The output problem is templates

This is the half that catches people after they have already chosen a tool, because it is not a language capability at all. It is a WhatsApp platform rule.

On the WhatsApp Business Platform you must assign a template language code when you create a template. Meta is explicit that template strings and variables are not translated for you, so you are responsible for supplying the strings and example parameters in their appropriate language. Templates are stored in language packs, one bundle per language or locale you support.

Practically, that means:

What you want	What you actually have to build
Order confirmation in English and Arabic	Two separate templates, each submitted and approved on its own
A campaign across the Gulf and the Levant	One template per language, plus per-market compliance checks
A new offer message	The Arabic version submitted for approval too, not translated at send time

Meta does let you create multiple templates that share a name in different languages, which keeps this manageable rather than chaotic. But the approval queue is per template, so an Arabic version can be rejected while the English one sails through, and you will not find out until you try to send.

The operational consequence is simple. Every template you add is really two templates, and your Arabic set needs the same review discipline as your English set. If your Arabic templates are machine translated once and never read by a native speaker, your customers will notice long before your dashboard does.

Getting the register right

Even correct Arabic can be wrong Arabic for a brand.

Modern Standard Arabic is understood everywhere and belongs to nobody. It is right for formal notices, legal wording and anything a customer might screenshot. It is wrong for a friendly nudge about a cart, where it reads stiff and slightly official.

Dialect is warmer and lands better in conversation, but it localises you. Levantine phrasing in a Riyadh campaign is noticeable in the way a British ad reads in Texas. It is not offensive, it is just visibly not from here.

The workable pattern for most brands selling across MENA is to keep sent templates in clear, plain Modern Standard Arabic that no market will find strange, and let live conversation adapt to how the customer writes. If they open in Emirati dialect, meet them there. If they open in Arabizi, reply in Arabizi. That is what a human employee in the shop would do.

What this means for a Gulf or Levant brand

Two other things matter regionally and are easy to forget.

Meta charges per message delivered on a rate that depends on the template category and the country calling code of the recipient. A single Arabic campaign that spans Saudi Arabia, the UAE and Lebanon is spanning several different rate cards, so budget by market rather than by campaign. We break the pricing model down in what the WhatsApp Business API really costs.

And the free part of the channel is language-agnostic. When a customer messages you, a 24 hour customer service window opens, and inside it you can reply without a template at all. That is where dialect handling earns its money, and where a fast, natural Arabic reply converts. Everything outside that window is a template you had to get approved in advance.

What to do this week

Pull twenty real Arabic customer messages from your own WhatsApp and build the test set described above. Do not clean them up.
Count how many of your existing templates exist in Arabic. If the answer is none, that is your first project, not your Arabic chatbot.
Have a native speaker from your main market read your Arabic templates out loud. Anything that makes them pause is costing you.
Check Meta's supported template languages and template fundamentals before you plan a multi-market rollout.

Arabic support is not a feature you switch on. It is a test set, a second template library, and a decision about register. plum is the AI that sells on WhatsApp, and it works in Arabic and English natively because MENA brands are who it was built for. If you want to see the setup end to end, read how it works, or start with our guide to WhatsApp AI for Shopify stores in Lebanon and the Gulf.

Questions people also ask

Can AI actually understand Gulf and Levantine dialects, or only Modern Standard Arabic?

Modern systems handle dialect far better than they did, but capability varies a lot by vendor and by dialect, and no vendor's marketing page will tell you where its limits are. The only reliable answer comes from testing with your own customers' messages, in dialect and Arabizi, with the typos left in. Twenty real messages will tell you more than any demo.

Do I need separate WhatsApp templates for Arabic and English?

Yes. Meta requires you to assign a language code when creating a template and does not translate template strings or variables for you, so each language is a separate template you submit and get approved. You can give templates in different languages the same name, which keeps them organised, but each one goes through review independently and can be approved or rejected on its own.

What is Arabizi and does it matter for customer service?

Arabizi is Arabic written in Latin characters, using numerals for sounds with no Latin equivalent, as in "3andkon" or "kifak". It is extremely common in messaging across the Levant and the Gulf, particularly among younger customers. It matters because systems that only parse Arabic script or English words treat it as noise, which means a real share of your inbound messages get misunderstood.

Should our automated messages use dialect or Modern Standard Arabic?

For templates you send first, plain Modern Standard Arabic is the safer default, because it is understood across every market and does not sound out of place in any of them. Dialect is warmer but localises you, so Levantine phrasing reads as foreign in the Gulf. In live conversation, matching how the customer writes is better than either default.

Does replying in Arabic cost more than replying in English?

No. Meta's rates depend on the template category and the recipient's country calling code, not on the language of the message. What does change your costs is where your customers are, since a campaign spanning Saudi Arabia, the UAE and Lebanon crosses several rate cards. Replies inside the 24 hour customer service window are a different matter and are not template sends.

Don't take our word for it

Have a question? Ask plum.
See it for yourself

Most companies can only describe their product. You can message ours. Ask plum about availability, its benefits to you, and skip the FAQ rabbit hole while you are at it. See how it answers, how it sells, and how it sounds. Then we will build you your own.

Chat with plum

ask a question book a meeting